What is a specialist agent?

The instinct when building with LLMs is to stuff everything into one agent — one prompt, every tool, unlimited scope. It works for demos. It breaks in production. The cure is specialization: several narrow, deep agents, each outstanding at its role, coordinated by a supervisor. This is the long-form definition of the atom. Short version at glossary/specialist-agent.

The precise definition

A specialist agent is an AI agent whose scope, prompt, tool set, context window, and evaluation criteria are all designed around a single role within a larger multi-agent system. The design principles are four: scope (do one thing), isolation (own context), parsimony (only the tools you need), and testability (clear rubric for the role's outputs).

In plain English

Hiring a senior content writer is different from hiring a generalist. The senior writer has a portfolio. She has a narrow set of tools she knows cold — a style guide, a CMS, a citation tracker. Her brief is bounded. Her work is judged against criteria that matter for writing (voice, clarity, accuracy) and not against criteria that don't (code quality, spreadsheet formulas).

A specialist agent is the software version. You design each one like you'd scope a role on a team: what's the job, what are the tools, what's the output, how do we know it's good. A coding agent doesn't need access to the email API. A content agent doesn't need a shell. An evaluator agent doesn't need to write anything — it needs to read and judge.

Specialization gives you three compounding advantages: each agent performs better on its role; each agent is easier to debug; and the whole system gets more reliable over time because you can improve one role at a time without regressing the others.

The history

The idea of role-based agents shows up early in multi-agent research — BDI architectures, FIPA-compliant service agents, KQML message types — but the LLM version came in waves. Anthropic's research on character and role prompting showed that persona instructions reliably shift behavior. Stanford's Generative Agents paper (2023) gave each agent a persona, a memory stream, and reflection. CAMEL (KAUST, 2023) demonstrated that role-playing agents solving the same task produced better results than identical copies.

On the engineering side, AutoGen's "AssistantAgent" and LangGraph's "node" abstractions formalized specialists as first-class citizens. CrewAI leaned hardest on the metaphor, with a "role / goal / backstory" tuple for every agent. The Claude Agent SDK packaged the same idea as subagents — named agents with their own prompts and tools, invoked by a parent agent.

Why specialists beat one big generalist

Three reasons, in order of importance:

1. Prompt quality. A focused prompt with 5 exemplars of great writing outperforms a mega-prompt that covers 20 capabilities. Working memory is finite even at million-token context; the model will attend more reliably to a 500-token role prompt than a 5,000-token one.
2. Context hygiene. When research, code, copy, and design all live in one agent's context, they pollute each other. A specialist's context stays clean. When it returns, only the output artifact crosses the boundary.
3. Evaluability. A coding agent is judged on tests passing, code diffs reviewed. A content agent is judged on voice match and structural checks. Two separate rubrics. One mega-agent has no clean rubric because the criteria conflict.

Anatomy of a specialist

A well-designed specialist has five parts:

• System prompt. Names the role, states the scope, lists the tools, shows 2-5 examples of great output, defines the return schema, and pins the acceptance criteria.
• Tool set. Three to seven tools, max. More tools means the model spends reasoning budget on selection instead of execution.
• Context. Fresh per invocation. Gets only what the supervisor passes in — no session-wide noise.
• Output schema. Structured output (JSON, markdown with a defined shape) so the supervisor can parse the result reliably.
• Rubric. What "good" looks like for this role, fed to the evaluator when its work is checked.

Why it is different from a subagent

"Subagent" is a structural term: any agent invoked by a parent agent is a subagent. "Specialist" is a design term: the subagent is scoped to a specific role. In practice, almost every production subagent is a specialist. But you can have a subagent that is a generalist clone — and in most cases, you should not.

Anti-patterns

• The kitchen-sink specialist. A "Business Ops" agent that also does research, writes copy, deploys code. Split it or accept that it won't be good at any of them.
• The naked persona. A system prompt that just says "You are a helpful assistant with expertise in marketing." No scope, no tools defined, no output schema. The model will improvise and you'll get inconsistent results.
• Peer-to-peer specialist calls. Specialist A calling Specialist B directly, bypassing the supervisor. Creates loops and unreadable audit trails.
• Over-specialization. Fifty specialists where fifteen would do. The supervisor can't route well, and the ongoing maintenance cost dominates the marginal quality gain.

Real-world example

A specialist agent for "Research," from a real Black Box deployment:

Role: Research specialist.
Scope: Given a topic and a desired output shape,
       gather credible sources, synthesize findings,
       return a structured brief. You do not write
       finished content; you produce briefs for the
       content specialist.

Tools: web_search, fetch_url, save_source_note.

Return shape: topic, key_findings (array), sources
              (array), confidence (float).

Acceptance criteria:
  - 3+ credible sources (prefer primary)
  - Key findings each link to a source
  - Confidence flag honest about contradictions

Examples: three briefs the team rates as 9/10.

When the CEO needs research, it sends the Research specialist a topic + shape. Research returns a brief. Everything stays scoped, parsable, and auditable.

How Black Box implements specialists

Black Box ships 18 specialists at Pro tier and above: Coding, Content, Research, Browser, Business Ops, Sales, Design, Analytics, Evaluator, and more. Each has a constitution file — system prompt, tool manifest, output schema, rubric — that lives in version control. The CEO agent invokes them through the Claude Agent SDK subagent primitive. Each specialist runs in its own context window. The features page shows the full roster; tiered access is on pricing.

The roster is not fixed. Skill Packs — our distribution unit for new capabilities — add new specialists or extend existing ones. See the post on Skill Packs for how.

Key takeaways

• A specialist agent is an AI agent scoped to one role with a matching prompt, tool set, context, and rubric.
• Specialists outperform generalists on their role and make the system debuggable.
• Anatomy: prompt, tools, context, output schema, rubric — in that order.
• Common anti-patterns: kitchen-sink, naked persona, peer-to-peer calls, over-specialization.
• Black Box ships 18 specialists plus a CEO, extensible via Skill Packs.

Frequently asked questions

Related reading

• Glossary: Specialist Agent
• What is an AI executive team?
• What is multi-agent orchestration?
• What are Skill Packs?